Enhanced Multi-Objective Design Optimisation of Salient Pole Reluctance Magnetic Gear Using Bayesian-Optimised Artificial Neural Networks

Abstract

The application of artificial intelligence in magnetic gear design has opened new avenues for accelerating computation and optimisation processes. In this paper, a Bayesian-optimised artificial neural network (ANN) was presented as a surrogate model to predict the performance of salient pole reluctance magnetic gears (SP-RMGs). The model focuses on key performance indicators such as average torque, torque ripple, and total weight. A diverse dataset generated through Latin hypercube sampling (LHS) is used to train the ANN, which employs customised activation functions to accurately capture the non-linear characteristics of the magnetic gear. Bayesian optimisation is applied to fine-tune the hyperparameters, resulting in a significant reduction in computational time. The proposed approach leverages deep learning to efficiently accelerate the multi-objective optimisation process, providing accurate predictions of SP-RMG performance metrics. The optimisation results demonstrate significant improvements with the model predicting optimal design parameters that enhance torque performance, reduce torque ripple by 47.2%, and decrease total weight. The proposed approach offers a substantial reduction in computational time while delivering precise optimisation outcomes.